Inside Project Marlin: How 1,000 Contractors Are Training Anthropic’s $2.5B Coding Agent

Anthropic’s Claude Code became the most dominant AI coding agent on the market by writing code better than its rivals—but a newly revealed initiative shows it didn't get there by machine learning alone. The effort, known internally at data vendor Snorkel AI as "Project Marlin," enlists roughly 1,000 contract software engineers to manually stress-test, evaluate, and fine-tune Claude Code's output, paying top dollar for their judgment .

The project, first reported by Business Insider in late May 2026, sheds light on a fast-evolving corner of the AI industry where the labor required to improve frontier models has shifted from simple data labeling to highly technical, expert-level feedback .

What Project Marlin Does

Project Marlin tasks contractors with creating realistic coding prompts and reviewing Claude Code's responses on real-world GitHub pull request scenarios, including feature additions and bug fixes . The work is structured as blind A/B tests: contractors compare outputs from two different model versions, pick the preferred code, and provide qualitative feedback, without knowing which model they are evaluating .

The scale and compensation reflect how much Anthropic values this expert input. Two contractors told Business Insider they were paid up to $280 per task, with each task typically taking about an hour, allowing some to earn over $3,000 per week . Snorkel AI maintains an internal approval layer to ensure quality control on submissions .

Why a $2.5 Billion Product Still Needs Human Judges

The scale of the investment in Project Marlin becomes clearer when examined against Claude Code's staggering commercial trajectory. The coding agent, which launched publicly in May 2025, hit a $1 billion annualized revenue run-rate by November of that year and doubled to $2.5 billion by February 2026 .

By the time details of Project Marlin emerged, Claude Code had overtaken Cursor and GitHub Copilot in terms of revenue, capturing an estimated 51% to 54% of the AI coding market . This ascent was fueled by a tool that Anthropic's own internal teams had come to rely on for 70% to 90% of their code, including approximately 90% of Claude Code's own codebase .

The Marlin initiative reveals a key insight about this success: even the most powerful AI coding agents still require sophisticated human feedback to close the gap between writing functional code and mimicking the nuanced judgment of a professional developer . The project's explicit goal is to fine-tune Claude Code to better replicate professional-level skills, moving beyond simple syntax correctness toward architecture decisions, code review sensibility, and contextual problem-solving .

The Industry Shift: From Data Labeling to Expert Evaluation

Project Marlin represents an important evolution in how AI companies approach training labor, particularly as coding agents become the highest-value enterprise use case for generative AI, accounting for 51% of all enterprise usage .

Traditional data-labeling workflows, where lower-cost workers annotate images or classify text, are ill-suited to evaluating a tool designed to reason through complex pull requests. Instead, companies like Anthropic are paying substantial premiums for contractors who can exercise engineering judgment, a trend that may accelerate as the economic stakes for AI coding tools continue to rise.

The broader implications for the labor market are significant: as AI models become more capable, the human oversight required to improve them does not disappear—it shifts upward in skill level and compensation. Project Marlin suggests the future of AI training may look less like a factory floor and more like an elite code review process, where top engineers are paid hourly to teach machines how to think like senior developers.